But idk how much the multispeaker architecture helps decent synthesis results, and if itd just be better to swap to a newer architecture for more accurate, coherent outputs

I mean to me a faster Tacotron2 and hifigan implementation seem like the way to go. I just personally don't know how to run them. And if you train your model on a wide range of speech with good sentence structure or pre processing of text you might limit stutter?

What's the min amount of training data you need for good output?

@User @User so that token

stuff yall were talking about

can I use that for Uberduck?

Or is that more 'beyond uberduck'type stuff?

woo mafia

You could try training it and see if itd work without modifying the model

Pretty good, what architecture is this? And what loss are you getting from that generation?

huh. yeah fair enough

"i will find him, i will capture him, and no one will ever die again" -spy

talknet, uhh heck didn't check

unguided it's abnormally high-pitch here

Did you train the vocoder too?

i think yeah

Hmm

i'm just using the notebooks like a baby

I'd compare the synthesizer output vs. Vocoder output

My instincts say the vocoder needs a lil more time

i know for a fact that it's not optimal, because i assumed talknet was letter-based so i did

thiiiiiis

in my training data

but that's probably a small factor

idk

That's fine,I'm assuming the models tokenize the input strings

yeah this one does arpabet

But if theres not a lot of other data like that in the pretrained dataset you can get lower results as it needs to learn from scratch

Ah

Right, I remember reading the paper

Yeah itll tokenize it into arpabet

Can you download one of the spectogram truth outputs from the GTA synthesis and play it?